If the tree is grown without fixing the depth, each leaf in the tree will be a sample. So you clearly overfit but you will do a perfect classification if you train and test on the same data. Basically, this is a behaviour that we want to check (I don't know if it was intended here).

it would be the same behaviour for regression.

If the tree is grown without fixing the depth, each leaf in the tree will be a sample.

Thanks, that makes sense.

I think the test copied from (or at least is the same as) this one in test_forest:

In test_forest the regressors are not completely overfit as each tree uses only a subset of the data and the scores are not perfect.

I'm not sure what this test is checking for - but if is checking that using less features reduces the learning ability (as suggested by the comment and the worse score value in the 2nd assert) then, it isn't doing a good job. We should restrict depth.

ping @glemaitre

This is kinda funny that we only have a single build failing here. The splits are probably different in 32 bits. The max_depth is probably not sufficient since the splits are completely random. You can increase the depth of the trees then.

But we set random state? For me(64 bit) overfitting occurs quickly. For max_depth=20, 4/6 scores are 0

DecisionTreeRegressor mse score 0.0
DecisionTreeRegressor mae score 5.786199095022624
DecisionTreeRegressor friedman_mse score 0.0
ExtraTreeRegressor mse score 0.0
ExtraTreeRegressor mae score 3.6911764705882355
ExtraTreeRegressor friedman_mse score 0.0

The score is larger than 60 with 32bit

Tree       = <class 'sklearn.tree._classes.ExtraTreeRegressor'>
criterion  = 'mse'
max_depth  = 15
name       = 'ExtraTreeRegressor'
reg        = ExtraTreeRegressor(max_depth=15, max_features=6, random_state=0)
score      = 281.43585091379208

The large difference between 32 and 64 is odd.

The large difference between 32 and 64 is odd.

Basically, I recall that the trees built with 32 bits architecture are different from the one built with 64 bits. So either we increase the threshold or we skip the test on 32 bits architecture with the decorator @skip_if_32bit from sklearn.utils._testing. Let's do the latest and see what people think about it in a second review.

Thank you @lucyleeow !